Angelos Mavrogiannis

I am a PhD student in the department of Computer Science at the University of Maryland, College Park. I am a member of the Perception and Robotics Group where I work in the intersection of Robotics and Natural Language Processing, advised by Prof. Yiannis Aloimonos.

Prior to this, I received a Master of Science in Mechanical Engineering from Carnegie Mellon University. At CMU, I was a member of the Intelligent Control Lab in the Robotics Institute and my thesis focused on driver behavior classification, advised by Prof. Changliu Liu. Prior to CMU, I received my diploma (BS & MEng) in Mechanical Engineering and Aeronautics from the University of Patras (Greece). For my diploma thesis, I worked with Prof. Argyris Dentsoras on the development of a framework for automatic parsing and solution of optimization problems formulated in structured language.

Email  /  CV  /  LinkedIn  /  Google Scholar  /  Twitter  /  Github  /  SoundCloud

Research

I am interested in the intersection of Robotics and Natural Language Processing (RoboNLP), and more specifically in leveraging linguistic traits to enrich robots with commonsense reasoning.

News

November 2024: I passed my PhD preliminary exam and advanced to candidacy. Here is my proposal document.
September 2024: I will be presenting recent work and motivating modeling Actions as Programs in the Modern Era at ICRA@40 in Rotterdam, Netherlands.
April 2024: Patent application filed for work I was involved in at LinkedIn during my summer 2023 internship.
March 2024: Cook2LTL was accepted for publication at IEEE ICRA 2024 and will be presented in Yokohama, Japan in May 2024.
November 2023: I am returning to LinkedIn for a summer 2024 internship in San Francisco.
March 2024: I will be attending The Machine Learning Summer School (MLSS) 2024 at the Okinawa Institute of Science and Technology (OIST) in Okinawa, Japan where I will present a poster on Cook2LTL.

We present a perception-action programming API that consists of Vision Language Models (VLMs) and Large Language Models (LLMs) as backbones, together with a set of robot control functions. When prompted with this API and a natural language query, an LLM generates a program to actively identify attributes given an input image.

Cook2LTL is a system that receives a cooking recipe in natural language form, reduces high-level cooking actions to robot-executable primitive actions through the use of LLMs, and produces unambiguous task specifications written in the form of LTL formulae.

B-GAP is a new simulation technique consisting of enriching existing traffic simulators with behavior-rich trajectories corresponding to varying levels of aggressiveness. After generating these trajectories with the help of a driver behavior modeling algorithm, we use an enriched simulator to train a Deep Reinforcement Learning (DRL) policy that consists of a set of high-level vehicle control commands and apply this policy at test time to perform local navigation in dense traffic.

We extract high-level features from raw vehicle trajectory data and classify drivers into behavioral classes based on their level of aggressiveness. We demonstrate how the identification of a driver's behavior improves the accuracy of the short-term trajectory prediction problem by introducing a prior knowledge on their behavior.

We developed a framework on Visual Basic that receives mathematical expressions as input, analyzes them using a suitable parser, and optimizes them with genetic algorithms. The parser allows the input of the expressions in string format and distinguishes the variables, the parameters and the operational symbols. Besides the equations, the user can choose between a set of genetic algorithms for the optimization, as well as the hyperparameters. The implemented software was tested and validated on two applications: the minimization of the forces applied onto an object grasped by a robotic arm and the maximization of the stiffness of a cantilever beam.

LinkedIn Internship Projects

In the summers of 2023 and 2022 I worked as an Artificial Intelligence - Machine Learning Engineer in the Ads AI organization at LinkedIn in Mountain View, California.

My summer 2023 intern project introduced open-source generative AI in the current LinkedIn audience targeting pipeline. I used Low Rank Adaptation (LoRA) to fine-tune state-of-the-art open-source Large Language Models, quantitatively compared them and identified the best performing task-specific model. Fine-tuning this model, I outperformed the existing baseline used in production for automatic audience targeting in user-defined advertising campaigns.

I developed a DRL framework for multi-constraint automatic bidding towards optimizing the LinkedIn Ads platform. The framework was built upon OpenAI gym, using synthetic data to simulate a 2nd price auction real-time bidding environment. This project was the stepping stone for a DRL-based system that was subsequently used in production for automatic bidding at LinkedIn.

Course Projects

Representative group course projects at UMD and CMU. For more projects check my CV.

In this work we measure the fairness of diffusion-based face generation models powered by conversational input data in textual form. We design a pilot study where participants answer LLM-generated non-intrusive questions about themselves with the goal of implicitly extracting facial features to build an informed prompt for face generation.

Proposed and led a group project on building a 3D-printed chessboard-resembling racetrack and an interactive system that converts it to a VRsimulated racing environment. Implemented the computer vision module, tracking the position and orientation of the pieces using ArUco markers and mapping them to poses and ego-vehicle control commands in a simulated racetrack in Unity.

This work uses a Hierarchical Attention Network architecture to assess the potential suicide risk of reddit users based on their post history. My contribution was the extraction of post-level features based on users’ emotional states and the tuning of a Latent Dirichlet Allocation (LDA) model to retrieve meaningful subreddit clusters.

Collaborated with a team of students to design and manufacture an underwater penguin-inspired robot. The movement of the flippers was powered by an Arduino microcontroller controlling a ball-and-socket motion transmission mechanism.

Collected a synthetic dataset of manipulator postures and object poses in OpenAI Gym. Trained a multilayer perceptron in order to map changes in hand pose to object displacements. Modified the OpenAI Gym simulator to demonstrate the predicted object pose and validated the method on occluded object tracking problems.

Summer Schools

Design and source code from Jon Barron's website. Do not scrape the HTML from this page itself, as it includes analytics tags that you do not want on your own website — use the github code instead.